high performance computing on graphics processing units: hgpu.org

Commodity many-core hardware is now mainstream, driven in particular by the evolution of general purpose graphics programming units (GPGPUs), but parallel programming models are lagging behind in effectively exploiting the available application parallelism. There are two principal reasons. First, real-world applications often exhibit a rich composition of nested parallelism, whose statical extraction requires a set of (compiler) transformations that are tedious to do by hand and may be beyond the capability of the common user. Second, the best optimization strategy, with respect to what to parallelize and what to sequentialize, is often sensitive to the input dataset, and as such, it may require several code versions, maintained and optimized differently for different classes of datasets. This paper studies three such array-based applications from the financial domain, which are suitable for GPGPU execution. For each application, we (i) describe all available parallelism via nested map-reduce functional combinators, (ii) describe the invariants and code transformations that govern the main trade-offs of a rich, dataset-sensitive optimizations space, and (iii) report target CPU and GPGPU code together with an evaluation that demonstrates optimization trade-offs and other difficulties. Finally, we believe this work provides useful insight into the language constructs and compiler infrastructure capable of expressing and optimizing such applications, and we report in progress work in this direction.